# Deepgram

> Deepgram is a speech recognition and voice AI platform sold as a developer API, covering real-time and pre-recorded transcription in more than fifty languages with speaker diarization, smart formatting, keyterm prompting, redaction, and audio intelligence models for summarization, topic detection, sentiment, and intent; it is self-serve pay-as-you-go from roughly $0.0043 per minute with $200 in free credits, and it provides no dashboard, call library, scorecards, or CRM integration of its own.

- Category: Conversation Intelligence (https://saastracker.org/categories/conversation-intelligence)
- Website: https://deepgram.com
- Starting price: $0.0043 per minute (Nova-3 pre-recorded monolingual), with $200 in free credits
- Free plan: The $200 credit funds free usage across models; there is no perpetual free tier beyond the credit.
- Free trial: $200 in free credits on signup with no credit card required
- Founded: 2015, HQ: San Francisco, California, United States, Ownership: Venture-backed
- Profile last reviewed: 2026-08-22
- Canonical profile: https://saastracker.org/products/deepgram

## Overview

Deepgram is infrastructure, and a buyer browsing conversation intelligence tools should know that before reading further. It has no recording bot, no meeting library, no coaching dashboard, no scorecards, no talk-ratio report, no deal board, and no CRM connector. It is the speech engine that other people's products are built on, and if you are a sales manager looking for something to log into on Monday morning, Deepgram is not it. What earns it a place on a small-business list is that it is self-serve, transparently priced by the minute, and cheap enough that a two-person startup can run a real product on it.

The company's distinguishing obsession is latency and cost per minute. Founded in 2015 by three former University of Michigan physicists, Scott Stephenson, Adam Sypniewski, and Noah Shutty, Deepgram trained its own models rather than fine-tuning someone else's, and the payoff is streaming transcription under 300 milliseconds and per-minute prices that undercut most of the market. Nova-3 handles production transcription across more than fifty languages, Flux is the newer conversational model with turn detection and interruption handling built for voice agents, and industry-tuned variants exist for healthcare, legal, and finance.

Pricing is unusually generous at the entry point. Signup is self-serve, $200 in credits arrive with no credit card, and pay-as-you-go rates run $0.0043 per minute for Nova-3 pre-recorded monolingual, $0.0048 for Nova-3 streaming, and $0.0052 to $0.0078 for multilingual variants. Speaker diarization on pre-recorded audio is included rather than charged, which is a meaningful difference from AssemblyAI. Add-ons are small: redaction at $0.0020 per minute, entity detection at $0.0017, keyterm prompting at $0.0013. A Growth plan at $4,000 or more a year buys roughly 13 to 20 percent off through prepaid credits.

The company is now large and well funded, having raised a $130M Series C at a $1.3B valuation in January 2026 led by AVP with Twilio, ServiceNow Ventures, SAP, and Citi Ventures among strategic investors, bringing total funding to roughly $169M across five rounds and headcount to around 310. It also offers self-hosted deployment, which is the feature that lets a company with data residency or air-gap requirements use it at all, and which no packaged conversation intelligence tool on this list can match.

## How it works

1. For a recording you post the audio file or a URL to the pre-recorded endpoint, choose a model, and enable the features you want as query parameters. The response is JSON with word-level timings, punctuation and capitalization from smart formatting, and speaker labels from diarization, which on the pre-recorded path costs nothing extra.

2. For live audio you open a streaming websocket and receive interim and final transcripts as speech arrives, typically under 300 milliseconds behind the speaker. Flux adds conversational turn detection and interruption handling, which matters if you are building an agent that has to know when the human has finished talking rather than merely what they said.

3. Accuracy on domain vocabulary is handled through keyterm prompting and custom vocabulary injection rather than by retraining. Feeding your product names, competitor names, and industry acronyms into the request raises recall on exactly the terms a conversation intelligence application cares about, which is the cheapest accuracy improvement available.

4. Audio intelligence models sit on top of the transcript. Summarization, topic detection, sentiment analysis, and intent recognition are available as separate calls priced per thousand tokens rather than per minute. Anything beyond that, such as scoring a call against a rubric or tracking objections, you build yourself by sending the transcript to a model of your choosing and storing the result in your own database.

## Best for

Engineering teams building voice agents, call analysis, or transcription features into their own software, and technically capable small companies with volume, latency, or data residency requirements that packaged per-seat tools cannot meet.

## Not the right fit for

- Sales managers and revenue leaders; there is no interface at all, so the coaching workflow this category is normally bought for does not exist and would have to be written from scratch.
- Anyone who wants CRM write-back, deal risk scoring, pipeline forecasting, or call scorecards, none of which Deepgram has any concept of.
- Non-technical buyers of any kind. The $200 free credit and clean pricing page make it look accessible, but there is no path from signup to usable output that does not go through code.
- Teams that want the analysis layer supplied rather than assembled; Deepgram's audio intelligence is thinner than AssemblyAI's and considerably thinner than a conversation-native API such as Symbl.
- Buyers who need a predictable fixed monthly invoice, since usage pricing tracks volume and the only way to cap spend is to build the cap yourself.

## Features

### Speech recognition models

The core asset, trained in-house rather than fine-tuned from someone else's weights.

- **Nova-3 pre-recorded**: The production transcription workhorse at $0.0043 per minute monolingual and $0.0052 multilingual, which is among the lowest published rates in the market.
- **Nova-3 streaming**: Real-time transcription at $0.0048 per minute monolingual and $0.0058 multilingual, with latency typically under 300 milliseconds.
- **Flux conversational model**: A newer streaming model built for spoken dialogue, with turn detection and natural interruption handling, at $0.0065 per minute English and $0.0078 multilingual.
- **Whisper Large hosted**: OpenAI's Whisper Large available at $0.0048 per minute for teams who want that specific model without hosting it.
- **Industry-tuned models**: Variants tuned for healthcare, legal, and finance vocabulary, which materially improves accuracy on jargon-heavy conversations.
- **More than fifty languages**: Nova-3 covers 50-plus languages including Spanish, French, Mandarin, Japanese, Arabic, and Hindi; Flux currently covers around ten.

### Transcript quality features

The parameters that decide whether the output is usable for conversation analysis.

- **Speaker diarization included**: Who spoke when is included on pre-recorded transcription rather than charged as an add-on, which is a direct cost advantage over AssemblyAI where it is a separate line.
- **Smart formatting**: Automatic punctuation, capitalization, and paragraphing, without which a transcript is technically correct and practically unreadable.
- **Keyterm prompting**: Supplying critical terms raises keyword recall by up to 90 percent according to the vendor, at $0.0013 per minute on streaming, which is the cheapest way to make competitor and product-name tracking reliable.
- **Custom vocabulary injection**: Product names, acronyms, and internal jargon can be supplied per request rather than requiring a retrained model.
- **Filler word capture**: Ums and ahs are retained rather than silently cleaned, which is what makes delivery coaching metrics possible if you build them.
- **Numerals and multichannel audio**: Spoken numbers convert to digits, and multichannel transcription handles stereo call recordings where each party is on a separate track, the cleanest possible input for talk-ratio analysis.
- **Noise robustness**: Models are trained for real-world conditions rather than clean studio audio, which matters for phone calls and noisy sales floors.

### Audio intelligence and privacy

The analysis layer, thinner than a dedicated understanding API but present.

- **Summarization**: Transcript summaries billed per thousand tokens rather than per minute, adequate for a recap and not for a scored review.
- **Topic detection**: Automatic identification of what a conversation is about, the basis of grouping a call archive without writing a classifier.
- **Sentiment analysis**: Tone classification across a conversation, which is how a tone curve or a relationship-health signal gets built downstream.
- **Intent recognition**: Detection of what the speaker is trying to accomplish, more useful in support and voice-agent contexts than in sales coaching.
- **Redaction**: Automatic removal of sensitive and personal information at $0.0020 per minute on streaming, the primitive for keeping card numbers out of your transcript store.
- **Entity detection**: Extraction of names, organizations, amounts, and dates at $0.0017 per minute, the substrate for competitor and pricing trackers.

### Deployment, scale, and voice agents

The capabilities that make Deepgram viable where SaaS tools are not.

- **Self-hosted deployment**: Models can run in your own infrastructure, which is the only route on this entire list for an organization with air-gap or strict residency requirements, and something no packaged tool offers.
- **Text to speech**: Aura-1 at $0.0150 and Aura-2 at $0.030 per thousand characters, with Flux TTS at $0.045, so a full voice loop can be built on one vendor.
- **Sub-300 millisecond latency**: Low enough that a live agent or real-time coaching overlay feels responsive rather than laggy, which is the technical reason products choose Deepgram.
- **Custom model training**: Proprietary model training for edge-case audio, available to customers whose accuracy problem cannot be solved with keyterms.
- **HIPAA and enterprise compliance support**: Compliance arrangements including HIPAA are supported, with self-hosting available where a business associate arrangement is insufficient.
- **Growth plan prepaid discounts**: From $4,000 a year, prepaid credits reduce per-minute rates by roughly 13 to 20 percent without moving to a bespoke enterprise contract.

## Use cases

- **Startup building a voice agent**: The agent has to hear the caller, know when they have stopped talking, and respond fast enough that the pause does not feel broken. Outcome: Flux handles turn detection and interruption natively with sub-300 millisecond latency, and Aura text to speech closes the loop on the same bill and the same vendor relationship.
- **Product team adding call transcription to their app**: Customers want searchable call records inside the product, and every packaged conversation intelligence tool would mean sending customer audio to a third-party interface the customer never agreed to. Outcome: Nova-3 transcription with included diarization runs at under half a cent a minute inside the company's own infrastructure, and the transcript never leaves a system the customer already trusts.
- **Contact centre with volume and a per-seat problem**: Two hundred agents on a QA platform priced per seat is unaffordable, and only a sample of calls gets reviewed as a result. Outcome: Every call is transcribed and analysed at per-minute cost with redaction applied, so coverage goes from a sample to everything and the bill tracks minutes rather than headcount.
- **Regulated organization that cannot send audio to a SaaS vendor**: Healthcare or financial conversations cannot leave the network, which rules out every hosted conversation intelligence product on the market. Outcome: Self-hosted Deepgram runs the models inside the organization's own environment with industry-tuned vocabulary, making automated call analysis possible where it otherwise simply was not.

## Pricing

Usage-based pay-as-you-go priced per minute of audio by model, with most transcript-quality features included and analysis add-ons charged per minute or per thousand tokens. An annual Growth commitment buys prepaid discounts. No seats.

- **Pay-As-You-Go**: $200 credit, then per-minute rates usage. No credit card required to start; Nova-3 pre-recorded $0.0043 per minute monolingual; Nova-3 streaming $0.0048 per minute monolingual; Speaker diarization included on pre-recorded; No minimum commitment. $200 buys roughly 775 hours of Nova-3 pre-recorded transcription, which is an unusually large free allowance.
- **Growth**: $4,000+ per year in prepaid credits. Roughly 13 to 20 percent discount on published rates; Same models and features as pay-as-you-go; Prepaid credit structure rather than a bespoke contract; Self-serve rather than sales-negotiated.
- **Multilingual and conversational models**: $0.0052 to $0.0078 per minute. Nova-3 multilingual $0.0052 pre-recorded, $0.0058 streaming; Flux English $0.0065 per minute; Flux multilingual $0.0078 per minute; Whisper Large hosted at $0.0048 per minute.
- **Add-ons and text to speech**: $0.0013 to $0.045 per minute or per 1k characters. Redaction $0.0020 per minute streaming; Entity detection $0.0017 per minute; Keyterm prompting $0.0013 per minute streaming; Aura-1 TTS $0.0150 and Aura-2 $0.030 per 1k characters; Audio intelligence billed per 1k tokens.
- **Enterprise and self-hosted**: Custom quoted. Self-hosted deployment in your own infrastructure; Custom model training; HIPAA and enterprise compliance arrangements; Volume pricing beyond the Growth tier.

Billing notes:

- Speaker diarization is included on pre-recorded transcription rather than charged separately, which is a direct and often decisive cost advantage over AssemblyAI on archive processing.
- Audio intelligence features such as summarization, topic detection, sentiment, and intent are billed per thousand tokens rather than per minute, so long calls cost proportionally more than the per-minute rates suggest.
- The $200 free credit is more than four times what AssemblyAI offers and funds roughly 775 hours of pre-recorded transcription, so evaluation is effectively free at small-business scale.
- There is no seat concept, so the per-recorded-user versus per-team-member distinction that dominates this category does not apply; any number of people can view results at no extra cost.
- Growth at $4,000 or more a year is a prepaid credit commitment rather than a contract negotiation, which keeps the discount path self-serve.
- Usage pricing has no ceiling by default, so cost control is your responsibility to implement in the application layer.

Value assessment: Deepgram is the cheapest credible speech engine on this list per minute of audio, and the inclusion of diarization on pre-recorded transcription plus a $200 starting credit makes the effective entry cost near zero. For anyone building software, that is exceptional value. For anyone comparing it against a packaged tool as a way to save money, the arithmetic is a trap: you would be spending weeks of engineering to replace a $29 seat, and the resulting thing would have no library, no scorecards, and no dashboard. Judge it as infrastructure and it is excellent. Judge it as a conversation intelligence product and it scores zero, because it is not one.

## Strengths

- Among the lowest published per-minute transcription rates in the market, with speaker diarization included on pre-recorded audio rather than billed separately.
- A $200 free credit with no credit card, funding roughly 775 hours of transcription, which makes evaluation effectively free at small-business volumes.
- Sub-300 millisecond streaming latency and the Flux conversational model with turn detection, which is why voice agent builders choose it over general transcription APIs.
- Self-hosted deployment, the only option on this list for organizations that genuinely cannot send audio to a hosted vendor.
- Keyterm prompting and custom vocabulary raise recall on product and competitor names dramatically for a fraction of a cent per minute, which is exactly what conversation tracking needs.
- Very well capitalized and independent: roughly $169M raised including a $130M Series C at a $1.3B valuation in January 2026, with Twilio, SAP, ServiceNow Ventures, and Citi Ventures as strategic backers.

## Limitations

- No product interface whatsoever: no library, no dashboard, no scorecards, no coaching workflow, no deal view, no CRM integration.
- The audio intelligence layer is thinner than AssemblyAI's, with no auto chapters or audio-level PII redaction, and no built-in LLM gateway for custom scoring, so you bring your own model.
- Audio intelligence billed per thousand tokens rather than per minute makes cost harder to forecast than the clean per-minute transcription rates suggest.
- Flux, the model that makes conversational products work, covers only about ten languages against Nova-3's fifty-plus.
- Nothing in the product addresses recording consent, disclosure, or two-party consent law; those obligations sit entirely with the application you build.
- Usage pricing with no default ceiling means an integration bug or a traffic spike produces a bill with no automatic brake.

## Comparisons

- **Deepgram vs AssemblyAI**: The two direct competitors. Deepgram is cheaper per minute, includes diarization on pre-recorded audio, gives $200 in credits against $50, and offers self-hosting and sub-300ms latency. AssemblyAI supplies more analysis out of the box, including auto chapters and audio-level PII redaction, plus an LLM Gateway for custom scoring in the same account. Take Deepgram for cost, latency, and deployment control; take AssemblyAI when you want the understanding models handed to you.
- **Deepgram vs Symbl.ai**: Symbl.ai sells conversation-native endpoints, including a Call Score API priced per criteria per conversation and a real-time assist API, which is much closer to a finished coaching capability. Deepgram sells the transcript and expects you to build everything above it. Choose Symbl if you want scoring as an API call and can live with Invoca ownership; choose Deepgram for lower unit cost and total control.
- **Deepgram vs Grain**: Grain is a finished product with recording, a shared call library, coaching, and CRM sync at around $29 a seat, plus free viewer licenses. Deepgram is a per-minute API with no interface at all. For a small sales team the comparison resolves instantly in Grain's favour; Deepgram only wins when you are building software rather than coaching reps.
- **Deepgram vs CallRail**: CallRail is the packaged answer for a services business that wants inbound calls tracked, transcribed, and analysed, at $150 to $195 a month with no engineering. Deepgram is what you would use to build that yourself at a fraction of the marginal cost and a large multiple of the effort. Buy CallRail unless you have engineers and a requirement CallRail does not meet.
- **Deepgram vs Avoma**: Avoma is the full meeting lifecycle platform with scorecards, semantic trackers, dialer capture, and a revenue intelligence module, from $19 a seat plus $29 modules. Deepgram is the speech layer a vendor like that licenses or replicates. The only reason to compare them is the build-versus-buy question, and for a sales team without engineers the honest answer is buy Avoma.

## Implementation

- Setup time: Under an hour to a first transcript with the free credit and an API key. Months to anything a non-engineer would call a product, since storage, playback, search, scoring, and reporting are all yours to write.
- Learning curve: Low for a developer; the API surface is small and the documentation is good. Not applicable to anyone else, because there is no interface to learn.
- Onboarding: Fully self-serve through the Deepgram console with $200 in credits and no credit card. Self-hosted and custom model work involves the vendor's team.
- Migration: Switching speech APIs is comparatively cheap because the interface is a request and a JSON response; the accumulated cost lives in your own post-processing rather than in the vendor. Historical audio can be reprocessed in bulk at the same per-minute rate, and multichannel recordings give better diarization than mixed-down files, so store call audio per-channel if you can.

## Platform, API & security

- Platforms: REST API for pre-recorded audio, Streaming websocket API, Client SDKs, Self-hosted deployment, Text to speech API
- API: The product is the API: pre-recorded and streaming speech to text, audio intelligence endpoints, text to speech, and custom model training, all self-serve with published per-minute pricing.
- Compliance: SOC 2, GDPR, HIPAA support, Self-hosted deployment for stricter requirements
- Data residency: Self-hosted deployment allows full control of data location, which is the strongest residency answer on this list.
- SSO: API key authentication for self-serve accounts; SSO and directory integration are enterprise agreement concerns.
- Security notes: Redaction removes sensitive information from transcripts at $0.0020 per minute on streaming. HIPAA and enterprise compliance are supported, and self-hosting exists for organizations that cannot send audio to a third party at all. Recording consent and two-party consent obligations belong entirely to the application built on top.

## Support

- Channels: Email support, Developer support on paid usage, Enterprise support agreements, Sales-assisted onboarding for self-hosted
- Documentation: Comprehensive developer documentation at deepgram.com covering every model, feature flag, and deployment option, with SDKs and reference implementations.
- Community: Active developer community through Discord, GitHub, and a substantial technical blog covering speech model architecture.

## Company

- Founded: 2015
- Founders: Scott Stephenson, Adam Sypniewski, Noah Shutty
- Headquarters: San Francisco, California, United States
- Ownership: Venture-backed
- Employees: Approximately 310 (2026), up from around 164 in 2024
- Funding: Approximately $169M raised across five rounds, including a $130M Series C at a $1.3B valuation in January 2026 led by AVP.

Funding history:

- Seed and Series A (2016 to 2019): Part of the total. Y Combinator, Wing, and In-Q-Tel among early backers.
- Series B (2021): $25M. Led by Tiger Global with Madrona and Alkeon participating.
- Series C (2026): $130M. Led by AVP at a $1.3B valuation in January 2026, with Twilio, ServiceNow Ventures, SAP, and Citi Ventures as strategic investors.

Timeline:

- 2015: Founded by Scott Stephenson, Adam Sypniewski, and Noah Shutty, three former University of Michigan physicists, to build speech recognition models from scratch.
- 2016: Goes through Y Combinator and begins selling speech to text as a developer API priced by the minute.
- 2021: Raises a $25M Series B led by Tiger Global as demand for real-time transcription grows with remote work.
- 2023: Ships the Nova model family, pushing streaming latency below 300 milliseconds and expanding language coverage past fifty.
- 2025: Launches Aura text to speech and the Flux conversational model with turn detection, positioning around voice agents rather than transcription alone.
- 2026: Raises a $130M Series C at a $1.3B valuation led by AVP in January, with Twilio, SAP, ServiceNow Ventures, and Citi Ventures investing, and headcount reaching roughly 310.

## Integrations

REST API for pre-recorded audio, Streaming websocket API for real-time audio, Client SDKs across major languages, Text to speech via Aura and Flux TTS, Self-hosted deployment in your own infrastructure, Any application, telephony platform, or data store you connect it to yourself, No native meeting platform, CRM, or dashboard integrations

## FAQ

### What is Deepgram?

Deepgram is a speech recognition and voice AI platform sold as a developer API. It transcribes recorded and live audio in more than fifty languages with speaker diarization, smart formatting, keyterm prompting, and redaction, and offers audio intelligence models for summarization, topic detection, sentiment, and intent. It is infrastructure, not an application.

### Does Deepgram record my calls or give me a coaching dashboard?

No. There is no recording bot, no call library, no scorecards, no talk-ratio report, no deal board, and no CRM integration. Deepgram returns JSON. Everything a manager would use has to be built on top of it. If you want a tool you can log into and coach reps with, look at Colibri, Spiky, Avoma, or Grain.

### How much does Deepgram cost?

Signup includes $200 in credits with no credit card. After that, Nova-3 pre-recorded is $0.0043 per minute monolingual and $0.0052 multilingual, Nova-3 streaming is $0.0048 and $0.0058, Flux is $0.0065 English and $0.0078 multilingual, and hosted Whisper Large is $0.0048. Add-ons such as redaction, entity detection, and keyterm prompting run between $0.0013 and $0.0020 per minute.

### Is speaker diarization included?

On pre-recorded transcription, yes, along with smart formatting. That is a real cost advantage over AssemblyAI, which charges $0.02 an hour for diarization on async and $0.12 on streaming. Streaming diarization on Deepgram is priced separately, so check the current rate if your product is real-time.

### Is there a per-seat price?

No. Deepgram has no seat concept at all, so the usual conversation intelligence question of whether you pay per recorded user or per team member does not apply. You pay per minute of audio processed, and any number of people can look at the results without changing the bill.

### Can Deepgram run in my own infrastructure?

Yes, and this is one of its genuinely distinctive features. Self-hosted deployment lets the models run inside your own environment, which is the only viable route on this list for organizations with air-gap or strict data residency requirements. It is an enterprise arrangement rather than a self-serve option.

### What about recording consent and two-party consent law?

Deepgram does nothing about it, by design. It processes audio you send. Obtaining consent, announcing recording, and complying with two-party consent jurisdictions are obligations of whatever application you build. Redaction is available at $0.0020 per minute streaming to keep sensitive data out of the transcript, which is a data-handling control rather than a consent one.

### Deepgram or AssemblyAI?

Deepgram is cheaper per minute, includes diarization on pre-recorded audio, gives four times the free credit, has lower latency, and can be self-hosted. AssemblyAI supplies a richer understanding layer including auto chapters and audio-level PII redaction, plus an LLM Gateway for custom scoring in the same account. Pick Deepgram for cost, latency, and control; pick AssemblyAI for analysis supplied rather than built.

### How would I build talk ratio or keyword tracking on Deepgram?

Diarization plus word-level timings gives you talk-to-listen ratio and longest monologue with arithmetic you write yourself. Keyterm prompting plus entity detection gives you reliable competitor and product-name tracking. Scoring a call against a rubric requires sending the transcript to a language model of your choosing, since unlike AssemblyAI there is no built-in LLM gateway.

### How stable is Deepgram as a vendor?

Very. Founded in 2015 by three physicists from the University of Michigan, headquartered in San Francisco with roughly 310 employees, and funded to about $169M including a $130M Series C at a $1.3B valuation in January 2026 with Twilio, SAP, ServiceNow Ventures, and Citi Ventures among strategic investors.

### How many languages does it support?

Nova-3 covers more than fifty languages including Spanish, French, Mandarin, Japanese, Arabic, and Hindi, with dedicated multilingual variants. Flux, the conversational model with turn detection, currently covers around ten, so a multilingual real-time product may have to trade the conversational features for the language coverage.

## Editorial verdict

Deepgram is the price and latency leader among independent speech APIs, and the only option on this list that can run inside your own network. Diarization included on pre-recorded audio, $200 in starting credits, sub-300 millisecond streaming, keyterm prompting that actually fixes product-name recall, and a $1.3B-valuation balance sheet behind it make it a safe and cheap foundation for anything you are building. Understand the boundary before you buy: this is JSON, not a product. No dashboard, no library, no scorecards, no CRM. Choose it when you are building voice or call analysis into software, when volume makes per-seat pricing absurd, or when the audio genuinely cannot leave your network. If you are a sales manager with a coaching problem and no engineers, buy a packaged tool instead and do not let the half-cent-a-minute price tag start an argument.

---

Source: SaaSTracker (https://saastracker.org), an independent editorial project. This profile is compiled from public information, carries no peer reviews or paid placement, and was last reviewed 2026-08-22. Awards are judged on published criteria: https://saastracker.org/methodology